Launching 8 eval workers...
  GPU 0 → PID 8530 → /mlx/users/jiashuo.fan/playground/inference/logs/eval_gpu0.log
  GPU 1 → PID 8531 → /mlx/users/jiashuo.fan/playground/inference/logs/eval_gpu1.log
  GPU 2 → PID 8532 → /mlx/users/jiashuo.fan/playground/inference/logs/eval_gpu2.log
  GPU 3 → PID 8533 → /mlx/users/jiashuo.fan/playground/inference/logs/eval_gpu3.log
  GPU 4 → PID 8534 → /mlx/users/jiashuo.fan/playground/inference/logs/eval_gpu4.log
  GPU 5 → PID 8535 → /mlx/users/jiashuo.fan/playground/inference/logs/eval_gpu5.log
  GPU 6 → PID 8536 → /mlx/users/jiashuo.fan/playground/inference/logs/eval_gpu6.log
  GPU 7 → PID 8537 → /mlx/users/jiashuo.fan/playground/inference/logs/eval_gpu7.log

All 8 workers started. Watch progress:
  tail -f /mlx/users/jiashuo.fan/playground/inference/logs/eval_gpu*.log